feat(local-asr): 加本地 Qwen3-ASR 引擎（macOS 优先）

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "openless-all/app/src-tauri/vendor/qwen-asr"]
+	path = openless-all/app/src-tauri/vendor/qwen-asr
+	url = https://github.com/antirez/qwen-asr.git

openless-all/app/src-tauri/Cargo.toml

@@ -12,6 +12,8 @@ crate-type = ["staticlib", "cdylib", "rlib"]
 
 [build-dependencies]
 tauri-build = { version = "2", features = [] }
+# Compile antirez/qwen-asr C sources for the local ASR engine on macOS.
+cc = "1.1"
 
 [dependencies]
 tauri = { version = "2", features = ["macos-private-api", "tray-icon"] }
@@ -24,7 +26,7 @@ serde_json = "1"
 tokio = { version = "1", features = ["full"] }
 tokio-tungstenite = { version = "0.24", features = ["rustls-tls-native-roots"] }
 futures-util = "0.3"
-reqwest = { version = "0.12", default-features = false, features = ["json", "multipart", "rustls-tls", "stream"] }
+reqwest = { version = "0.12", default-features = false, features = ["json", "multipart", "rustls-tls", "native-tls", "stream"] }
 thiserror = "1"
 anyhow = "1"
 log = "0.4"
@@ -52,6 +54,8 @@ core-graphics = "0.24"
 objc2 = "0.5"
 objc2-foundation = "0.2"
 objc2-app-kit = "0.2"
+# antirez/qwen-asr 用 malloc 分配返回字符串，必须用 C `free()` 释放。
+libc = "0.2"
 
 [target.'cfg(target_os = "windows")'.dependencies]
 raw-window-handle = "0.6"

openless-all/app/src-tauri/build.rs

@@ -1,3 +1,56 @@
 fn main() {
+    #[cfg(target_os = "macos")]
+    build_qwen_asr_macos();
+
     tauri_build::build();
 }
+
+/// 编译 antirez/qwen-asr 的 C 源（仅 macOS）。
+///
+/// 上游 Makefile `make blas` 等价配置：BLAS 加速通过 Accelerate framework，
+/// `USE_BLAS` + `ACCELERATE_NEW_LAPACK` 是必要宏。
+/// `-march=native` 这里**不**用——分发二进制要可移植，cc crate 在 release 下
+/// 默认带 `-O2`，加上 `-O3` 提一档；NEON/AVX 在源码里有 `#ifdef` 自动分派。
+#[cfg(target_os = "macos")]
+fn build_qwen_asr_macos() {
+    const VENDOR: &str = "vendor/qwen-asr";
+    const SOURCES: &[&str] = &[
+        "qwen_asr.c",
+        "qwen_asr_kernels.c",
+        "qwen_asr_kernels_generic.c",
+        "qwen_asr_kernels_neon.c",
+        "qwen_asr_kernels_avx.c",
+        "qwen_asr_audio.c",
+        "qwen_asr_encoder.c",
+        "qwen_asr_decoder.c",
+        "qwen_asr_tokenizer.c",
+        "qwen_asr_safetensors.c",
+    ];
+
+    let mut build = cc::Build::new();
+    build
+        .include(VENDOR)
+        .define("USE_BLAS", None)
+        .define("ACCELERATE_NEW_LAPACK", None)
+        .flag("-O3")
+        .flag("-ffast-math")
+        // 上游开 `-Wall -Wextra`；我们把 antirez 的代码当三方依赖，把无关警告压成静默
+        // 避免 build log 噪音淹没我们自己的告警。
+        .flag("-Wno-unused-parameter")
+        .flag("-Wno-unused-variable")
+        .flag("-Wno-unused-function")
+        .flag("-Wno-sign-compare")
+        .warnings(false);
+
+    for src in SOURCES {
+        let path = format!("{}/{}", VENDOR, src);
+        println!("cargo:rerun-if-changed={}", path);
+        build.file(path);
+    }
+    println!("cargo:rerun-if-changed={}/qwen_asr.h", VENDOR);
+
+    build.compile("qwen_asr");
+
+    // BLAS = Accelerate
+    println!("cargo:rustc-link-lib=framework=Accelerate");
+}

openless-all/app/src-tauri/src/asr/local/cache.rs

@@ -0,0 +1,132 @@
+//! 本地 Qwen3-ASR 引擎缓存。
+//!
+//! 用途：避免每次 dictation 都重加载 1.2GB+ 模型。引擎一次 load 后驻留在内存，
+//! 跨多次会话复用；用户在设置里决定"说完话即释放" / "保持 N 秒后释放" /
+//! "不释放"。
+//!
+//! 调度规则：每次会话结束后 spawn 一个 sleep+check 任务；任务在到点时检查
+//! `last_used`——如果中间又被使用过则不释放，否则 drop 引擎让 OS 回收 RAM。
+
+use std::path::Path;
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use anyhow::Result;
+use parking_lot::Mutex;
+
+#[cfg(target_os = "macos")]
+use super::QwenAsrEngine;
+
+pub struct LocalAsrCache {
+    #[cfg(target_os = "macos")]
+    inner: Mutex<Option<CachedEngine>>,
+    #[cfg(not(target_os = "macos"))]
+    _phantom: (),
+}
+
+#[cfg(target_os = "macos")]
+struct CachedEngine {
+    model_id: String,
+    engine: Arc<QwenAsrEngine>,
+    last_used: Instant,
+}
+
+impl Default for LocalAsrCache {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl LocalAsrCache {
+    pub fn new() -> Self {
+        Self {
+            #[cfg(target_os = "macos")]
+            inner: Mutex::new(None),
+            #[cfg(not(target_os = "macos"))]
+            _phantom: (),
+        }
+    }
+
+    /// 取已缓存的同 id 引擎，没有就加载（**阻塞、可能数秒**——调用方应放
+    /// `spawn_blocking`）。模型 id 不同则把旧的 drop 再加载新的。
+    #[cfg(target_os = "macos")]
+    pub fn get_or_load(&self, model_id: &str, model_dir: &Path) -> Result<Arc<QwenAsrEngine>> {
+        {
+            let mut slot = self.inner.lock();
+            if let Some(cached) = slot.as_mut() {
+                if cached.model_id == model_id {
+                    cached.last_used = Instant::now();
+                    log::info!("[local-asr cache] reuse engine: {model_id}");
+                    return Ok(Arc::clone(&cached.engine));
+                }
+                log::info!(
+                    "[local-asr cache] active model changed {} -> {}, drop old",
+                    cached.model_id,
+                    model_id
+                );
+                slot.take();
+            }
+        }
+        log::info!("[local-asr cache] loading {model_id} from {}", model_dir.display());
+        let engine = Arc::new(QwenAsrEngine::load(model_dir)?);
+        let mut slot = self.inner.lock();
+        *slot = Some(CachedEngine {
+            model_id: model_id.to_string(),
+            engine: Arc::clone(&engine),
+            last_used: Instant::now(),
+        });
+        log::info!("[local-asr cache] loaded {model_id}");
+        Ok(engine)
+    }
+
+    /// 标记最近使用时间——end_session 在调过 transcribe 之后调一下，
+    /// 让 release 计时器从这一刻重新算。
+    pub fn touch(&self) {
+        #[cfg(target_os = "macos")]
+        {
+            if let Some(cached) = self.inner.lock().as_mut() {
+                cached.last_used = Instant::now();
+            }
+        }
+    }
+
+    /// 如果空闲时长 ≥ threshold，释放引擎。返回是否真释放了。
+    pub fn release_if_idle(&self, idle_threshold: Duration) -> bool {
+        #[cfg(target_os = "macos")]
+        {
+            let mut slot = self.inner.lock();
+            if let Some(cached) = slot.as_ref() {
+                if cached.last_used.elapsed() >= idle_threshold {
+                    log::info!(
+                        "[local-asr cache] release engine {} after idle {:?}",
+                        cached.model_id,
+                        cached.last_used.elapsed()
+                    );
+                    slot.take();
+                    return true;
+                }
+            }
+        }
+        let _ = idle_threshold;
+        false
+    }
+
+    /// 立刻释放（用户点"立即释放"或删模型时调）。
+    pub fn release_now(&self) {
+        #[cfg(target_os = "macos")]
+        {
+            if let Some(cached) = self.inner.lock().take() {
+                log::info!("[local-asr cache] release engine {} on demand", cached.model_id);
+            }
+        }
+    }
+
+    pub fn loaded_model_id(&self) -> Option<String> {
+        #[cfg(target_os = "macos")]
+        {
+            return self.inner.lock().as_ref().map(|c| c.model_id.clone());
+        }
+        #[cfg(not(target_os = "macos"))]
+        None
+    }
+}